slides minimax problems - hkust€¦ · form as it is an lp and we know that lps do not have...

34
Minimax Problems Daniel P. Palomar Hong Kong University of Science and Technolgy (HKUST) ELEC547 - Convex Optimization Fall 2009-10, HKUST, Hong Kong

Upload: letu

Post on 01-May-2018

214 views

Category:

Documents


1 download

TRANSCRIPT

Minimax Problems

Daniel P. Palomar

Hong Kong University of Science and Technolgy (HKUST)

ELEC547 - Convex Optimization

Fall 2009-10, HKUST, Hong Kong

Outline of Lecture

• Introduction

• Matrix games

• Bilinear problems

• Robust LP

• Convex-concave games

• Lagrange game

• Summary

Daniel P. Palomar 1

Introduction

• Games are optimization problems with more than one decision

maker (or player, using the terminology of game theory), often

with conflicting goals.

• The general theory of games is complicated, but some classes of

games are closely related to convex optimization and have a nice

theory.

• Game theory goes back to the mid-1990s (1947 book by von

Neumann and Morgenstein and Nash theorem in 1950).

Daniel P. Palomar 2

Foreword by Nash in 2008 JSAC Special Issue

Daniel P. Palomar 3

Matrix Game

• A matrix game is a two-player zero-sum discrete game that can be

described by a payoff matrix.

• Player 1 chooses his strategy (index k) to minimize the payoff Pkl.

• Player 2 chooses his strategy (index l) to maximize the payoff Pkl.

• Pure strategies: each user is allowed to choose only one strategy.

• Mixed strategies: each user is allowed to randomize over a set of

strategies.

Daniel P. Palomar 4

Pure-Strategy Matrix Game

• Player 1 can expect that, for a given choice k, player 2 will choose

the largest entry in row k. This means that he has to choose the

row with the minimum largest entry:

mink

maxl

Pkl.

• Similarly, player 2 expects that, for a given choice l, player 1 will

choose the smallest entry in column l. He will then choose the

column with the maximum smalles entry:

maxl

mink

Pkl.

• But, which one is the correct formulation? Are they the same?

Daniel P. Palomar 5

Pure-Strategy Matrix Game (II)

• In general, the two formulations lead to different results.

• However, we can show that one is smaller than the other.

Lemma:

maxl

mink

Pkl ≤ mink

maxl

Pkl.

Proof.min

eyf (x, y) ≤ max

exf (x, y) ∀x, y

maxex

miney

f (x, y) ≤ miney

maxex

f (x, y) .

Daniel P. Palomar 6

Pure-Strategy Matrix Game (II)

• But, in general, for pure-strategy matrix games

maxl

mink

Pkl 6= mink

maxl

Pkl.

• This is related to the concept of Nash equilibrium (NE).

• If the game admits a NE, then

maxl

mink

Pkl = mink

maxl

Pkl.

• Nash famous theorem shows that for games with N players, under

some conditions, a NE exists.

Daniel P. Palomar 7

Mixed-Strategy Matrix Game

• Now consider that each player chooses a strategy with some

probability (independent of the other user).

• Suppose that u and v are the mixed strategies for players 1 and 2.

• The expected payoff is∑

i,j

uivjPij = uTPv.

• Reasoning as before, player 1 will choose u as the solution to

minu

maxv

uTPv

s.t. u ≥ 0 , 1Tu = 1

v ≥ 0 , 1Tv = 1.

Daniel P. Palomar 8

Mixed-Strategy Matrix Game (II)

• Since maxv

uTPv = maxi

[PTu

]i, we can rewrite the problem as

minu

maxi

[PTu

]i

s.t. u ≥ 0 , 1Tu = 1

or, in epigraph form,

(P1)

minimizeu,t

t

subject to u ≥ 0 , 1Tu = 1

t1 ≥ PTu.

Daniel P. Palomar 9

Mixed-Strategy Matrix Game (III)

• Similarly, for player 2 we have

(P2)

maximizeu,µ

µ

subject to v ≥ 0 , 1Tv = 1

Pv ≥ µ1.

• Since max-min ≤ min-max,

p⋆2≤ p⋆

1.

• We can interpret the difference p⋆1− p⋆

2as the advantage conferred

on a player by knowing the opponent’s strategy.

Daniel P. Palomar 10

Mixed-Strategy Matrix Game (IV)

• In this case, however, we have a stronger characterization:

Theorem: For feasible mixed-strategy matrix games:

p⋆1

= p⋆2.

Proof. Let’s find the dual of problem P1. The Lagrangian is

L (t,u; λ, σ, µ) = t + λT(PTu− t1

)− σTu− µ

(1Tu− 1

)

= t(1 − λT1

)+ (Pλ − σ − µ1)T

u + µ,

Daniel P. Palomar 11

Mixed-Strategy Matrix Game (V)

Proof. (cont’d) the dual function is

g (λ, σ, µ) =

{µ if λT1 = 1 and Pλ = σ + µ1 ≥ µ1

−∞ otherwise,

and the dual problem is

maximizeλ,µ

µ

subject to λ ≥ 0 , 1Tλ = 1Pλ ≥ µ1

which happens to be problem P2. The result follows because problems P1 and P2

are dual of each other and strong duality holds (since the LPs are feasible).

Daniel P. Palomar 12

Bilinear Problems

• Consider the following bilinear problem (more general than the

mixed-strategy matrix game):

minx

maxy

xTPy

s.t. Ax ≤ b

Cy ≤ d.

• Again, max-min ≤ min-max, but do we have equality?

• In this case, we cannot proceed as before since maxy

xTPy 6=

maxi

[PTx

]i.

Daniel P. Palomar 13

Bilinear Problems (I)

• Sometimes, depending on the problem structure, the inner maxi-

mization can be solved in closed form (like in the application of

worst-case robust beamforming design).

• In general, however, the inner maximization does not have a closed

form as it is an LP and we know that LPs do not have closed-form

solutions (except in particular cases).

• We can resort to duality to deal with the inner maximization

maxy

xTPy

s.t. Cy ≤ d .

Daniel P. Palomar 14

Bilinear Problems (II)

Lemma: The dual problem of the inner maximization is

minλ

dTλ

s.t. PTx = cTλ

λ ≥ 0 .

Proof. The Lagrangian is

L (y;λ) =(PTx

)Ty + λT (d− Cy)

=(PTx − CTλ

)Ty + dTλ,

the dual function isg (λ) =

{dTλ if PTx = CTλ

−∞ otherwise,

and the dual problem follows.

Daniel P. Palomar 15

Bilinear Problems (III)

• We are now ready to rewrite the minimax problem in the more

convenient form of a minimization problem:

Lemma: The original minimax problem (assuming feasibility of the

inner maximization) can be rewritten as the following LP:

(Minimax-LP)

minimizex,λ

dTλ

subject to PTx = CTλ

λ ≥ 0

Ax ≤ b.

Proof. Since the original problem has a nonempty feasible set, strong duality

holds and the primal value is equal to the dual value.

Daniel P. Palomar 16

Bilinear Problems (IV)

• Let’s deal now with the maximin formulation (as opposed to the

minimax):

maxy

minx

xTPy

s.t. Ax ≤ b

Cy ≤ d.

Lemma: The dual problem of the inner minimization is

maxν

−bTν

s.t. Py + ATν = 0

ν ≥ 0 .

Daniel P. Palomar 17

Bilinear Problems (V)

Lemma: The original maximin problem (assuming feasibility of the

inner minimization) can be rewritten as the following LP:

(Maximin-LP)

maximizey,ν

−bTν

subject to Py + ATν = 0

ν ≥ 0

Cy ≤ d.

• So far, we have been able to rewrite the maximin and minimax

problems as LPs.

Daniel P. Palomar 18

Bilinear Problems (VI)

• At this point, however, we don’t know yet if maximin equals

minimax, which is answered in the next result:

Theorem: For feasible bilinear problems:

max-min = min-max.

Proof. It follows by showing that (Minimax-LP) and (Maximin-LP) are dual ofeach other and strong duality for feasible LPs.

Let’s find the dual of (Minimax-LP). The Lagrangian is

L (x,λ; ν,σ,µ) = dTλ + νT (Ax− b) − σTλ + µT(PTx− CTλ

)

= (d− σ − Cµ)Tλ +

(ATν + Pµ

)Tx− bTν,

Daniel P. Palomar 19

Bilinear Problems (VII)

Proof. (cont’d) the dual function is

g (ν, σ, µ) =

{−bTν if d− Cµ = σ ≥ 0 and ATν + Pµ = 0

−∞ otherwise,

and the dual problem is

maximizeν,µ

−bTν

subject to Pµ + ATν = 0

ν ≥ 0

d ≥ Cµ

which is identical to (Maximin-LP).

Daniel P. Palomar 20

Robust LP

• An LP is of the form:

minimizex

cTx

subject to aTi x + bi ≤ 0 i = 1, · · · ,m.

• The parameters of the problem c, (ai, bi) for i = 1, · · · ,m are

usually assumed to be perfectly known.

• In real applications, however, there may be uncertainty in these

parameters. This leads to robust optimization.

• We will next consider the case where knowledge of the parameter

ai is imperfect: ai ∈ Ai where the uncertainty set Ai is known.

Daniel P. Palomar 21

Robust LP (I)

• The uncertainty set can be modeled in different ways, e.g., as an

ellipsoid (as in the application of robust beamforming) or as a

polyhedron.

• We will model Ai as an affine transformation of a polyhedron:

Ai = {ai = ai + Biui, Diui ≤ di} .

• The robust LP (to imperfect knowledge of ai) is then

minimizex

cTx

subject to supui∈Ui

(ai + Biui)Tx + bi ≤ 0 i = 1, · · · ,m

where Ui = {ui | Diui ≤ di}.

Daniel P. Palomar 22

Robust LP (II)

• We can now rewrite the robust formulation as

minimizex

cTx

subject to fi (x) ≤ 0 i = 1, · · · ,m

where fi (x) is the optimal value of

maximizeui

(xTBi

)ui +

(aT

i x + bi

)

subject to Diui ≤ di

whose dual problem is

minimizezi

dTi zi +

(aT

i x + bi

)

subject to DTi zi = BT

i x

zi ≥ 0.

Daniel P. Palomar 23

Robust LP (III)

• Now, from strong duality, the dual value equals the maximum value

fi (x). Therefore, we have the following result:

Theorem: The robust LP can be rewritten as the following LP:

minimizex,{zi}

cTx

subject to dTi zi +

(aT

i x + bi

)≤ 0 i = 1, · · · ,m

DTi zi = BT

i x

zi ≥ 0.

Daniel P. Palomar 24

Convex-Concave Games

• Consider now a more general game where the payoff function is

f (x,y) with player 1 minimizing over x and player 2 maximizing

over y.

• We already know that

maxy

minx

f (x,y) ≤ minx

maxy

f (x,y) .

• We say that (x⋆,y⋆) is a solution of the game (Nash equilibrium

or saddle point) if

f (x⋆,y) ≤ f (x⋆,y⋆) ≤ f (x,y⋆) ∀x,y.

• In words: at a saddle point, neither player can do better by

unilaterally changing his strategy.

Daniel P. Palomar 25

Convex-Concave Games (I)

• Convex-concave problems have been extensively characterized and

we now state two of the nicest results:

Lemma: If f is convex-concave (and some other conditions), then a

saddle-point exists.

Lemma: If a saddle-point exists, then max-min = min-max.

Proof. From the existence of a saddle-point, we have

f (x⋆,y⋆) = minx

f (x,y⋆) ≤ maxy

minx

f (x,y)

as well asf (x⋆,y⋆) = max

yf (x⋆,y) ≥ min

xmax

yf (x⋆,y) ,

Daniel P. Palomar 26

Convex-Concave Games (II)

Proof. (cont’d) which implies

maxy

minx

f (x,y) ≥ minx

maxy

f (x⋆,y) .

This combined with the well-known inequality

maxy

minx

f (x,y) ≤ minx

maxy

f (x,y)

shows thatmax

ymin

xf (x,y) = min

xmax

yf (x,y) .

Daniel P. Palomar 27

Lagrange Game

• We will now study a very particular game related to Lagrange

duality.

• Consider the Lagrangian of a general optimization problem:

L (x;λ) = f0 (x) +

m∑

i=1

λifi (x) .

• Observe that

supλ≥0

L (x;λ) =

{f0 (x) if fi (x) ≤ 0 i = 1, · · · ,m

+∞ otherwise

Daniel P. Palomar 28

Lagrange Game (I)

• Therefore, we can express the optimal value of the primal problem

asp⋆ = inf

xsupλ≥0

L (x;λ) .

• On the other hand, by definition of dual function (as infimum of

the Lagrangian), we have

d⋆ = supλ≥0

g (λ) = supλ≥0

infx

L (x;λ) .

• Thus, from the fact that max-min ≤ min-max, we have

d⋆ ≤ p⋆,

i.e., weak duality holds!

Daniel P. Palomar 29

Lagrange Game (II)

• Recall that max-min = min-max for a convex-concave function.

• Now, if the primal problem is convex, then the Lagrangian L (x;λ)

is convex in x and concave (linear in fact) in λ.

• Thus (caution: additional assumptions are needed as the dual

feasible set is not compact),

d⋆ = p⋆,

i.e., strong duality holds.

Daniel P. Palomar 30

Summary

• We have explored minimax problems (two-player zero-sum games).

• In practice, minimax formulations appear in robust optimization.

• In some cases, a saddle-point or Nash equilibrium exists and then

max-min = min-max.

• Different ways to deal with minimax problems: 1) use specific

numerical methods for minimax problems, 2) solve in closed-form

the inner maxim., 3) rewrite the inner maxim. as the dual minimiz.,

4) use a technique like the S-lemma to get rid of the inner maxim.

• Game theory considers the more general and complicated case of

N players.

Daniel P. Palomar 31

References on Minimax

• R. T. Rockafellar, Convex Analysis , 2nd ed. Princeton, NJ: Princeton Univ. Press, 1970.

• D. P. Bertsekas, A. Nedic, and A. E. Ozdaglar, Convex Analysis and Optimization, Athena

Scientific, Belmont, MA, 2003.

• Daniel P. Palomar, John M. Cioffi, and Miguel A. Lagunas, “Uniform Power Allocation in

MIMO Channels: A Game-Theoretic Approach,” IEEE Trans. on Information Theory, vol.

49, no. 7, July 2003.

• Antonio Pascual-Iserte, Daniel P. Palomar, Ana Perez-Neira, Miguel A. Lagunas, “A Robust

Maximim Approach for MIMO Comm. with Imperfect CSI Based on Convex Optimization,”

IEEE Trans. on Signal Processing, vol. 54, no. 1, Jan. 2006.

• Jiaheng Wang and Daniel P. Palomar, “Worst-Case Robust Transmission in MIMO Channels

with Imperfect CSIT,” IEEE Trans. on Signal Processing , vol. 57, no. 8, pp. 3086-3100,

Aug. 2009.

Daniel P. Palomar 32

References on Game Theory

• J. Osborne and A. Rubinstein, A Course in Game Theory , MIT Press, 1994.

• Gesualdo Scutari, Daniel P. Palomar, and Sergio Barbarossa, “Competitive Design of Multiuser

MIMO Systems based on Game Theory: A Unified View,” IEEE JSAC: Special Issue on Game

Theory , vol. 25, no. 7, Sept. 2008.

• Gesualdo Scutari, Daniel P. Palomar, and Sergio Barbarossa, “Optimal Linear Precoding

Strategies for Wideband Noncooperative Systems Based on Game Theory,” IEEE Trans. on

Signal Processing , vol. 56, no. 3, March 2008.

• Gesualdo Scutari, Daniel P. Palomar, and Sergio Barbarossa, “Asynchronous Iterative Water-

Filling for Gaussian Frequency-Selective Interference Channels,” IEEE Trans. on Information

Theory , vol. 54, no. 7, July 2008.

• Gesualdo Scutari, Daniel P. Palomar, and Sergio Barbarossa, “Cognitive MIMO Radio: A

Competitive Optimality Design Based on Subspace Projections,” IEEE Signal Processing

Magazine, Nov. 2008.

Daniel P. Palomar 33